Customer retention

What does the retention curve say?

It says one thing, and it says it in the shape of the line rather than in any single number. Plot the share of each signup cohort still doing the thing that matters in month 1, month 2, month 3 and so on. If the line falls and keeps falling toward zero, you have a product people try. If the line falls and then goes flat, you have a product people keep, and the height of the flat part is the business.

The test is arithmetic, not judgment. Take the gaps between consecutive periods for one mature cohort. Say it reads 42% in month 3, 38% in month 4, 36% in month 5, 35% in month 6 and 35% in month 7. The gaps are 4, then 2, then 1, then 0. They are shrinking toward nothing, so the curve is flattening at around 35%. If the gaps are 9, 8, 8, 7, nothing is flattening and the only honest read is that you are renting users.

Two things make a flattening curve look like a failing one, and both are definition problems rather than product problems. The first is counting signups who never activated: a cohort that includes people who never reached the product's first value will always collapse between period 0 and period 1, and the collapse tells you about your signup form, not your retention. The second is averaging immature cohorts into the tail, which drags the far end of the line down because the newest cohorts have only one or two periods of data. How to read a retention curve walks the whole build in a spreadsheet, with both mistakes drawn out.

For a sense of scale, the most carefully assembled public set of ranges I know of is Lenny Rachitsky's, collected from 20 growth leads and published in June 2020. On 6-month user retention it puts "good" at around 25% for consumer social, around 40% for consumer SaaS, around 60% for SMB and mid-market SaaS and around 75% for enterprise SaaS, with "great" roughly 20 points above each. On the revenue side, ChartMogul's 2023 report across more than 2,100 SaaS businesses put top-quartile net revenue retention at 94% for companies between $1M and $3M of ARR, rising past 105% above $15M. Mobile is a different world: AppsFlyer data reported by Braze in July 2026 puts day 30 retention across app categories at roughly 1% to 15%.

Read those as the altitude of the room you are standing in, not as a target. The number that decides whether your retention is good enough is your own payback period, which belongs to the Revenue stage.

Which behavior predicts staying?

One action, usually inside the first week, separates people who are still there in 90 days from people who are not. Finding it takes an afternoon and does not need a data team.

Pull three months of signup cohorts, old enough that you know who survived to day 90. Split them in two: retained and gone. Then count, for the first seven days only, how many times each group did each of eight to ten countable actions. Invited a teammate. Connected their own data. Created a second project. Used the feature the product is named after. Opened the product on a second day. You are looking for the action with the widest gap between the two groups that enough people in the retained group actually did. Both halves of that sentence matter. An action with a 40-point gap that only 3% of retained users ever took is a curiosity. An action with a 20-point gap that 70% of retained users took is your activation event.

Then be honest about what you found, because this is the step where most teams waste a quarter. A gap is a correlation. People who add a profile photo may stay longer because engaged people add photos, not because photos cause engagement. The cheap way to tell the difference is to change the onboarding so that a chunk of new users are pushed toward the action, hold out the rest, and see whether the pushed group's day 90 retention moves. If it does not, you found a symptom. Keep it as a predictor for scoring accounts and stop trying to engineer it.

The method and its traps sit in more detail under the Activation stage, because the action lives in the first session even though it pays off here.

What should lifecycle messages do, and when?

Each message should answer one decision the customer is actually making at that moment, and it should be triggered by what they did or did not do rather than by the calendar. Most companies have one welcome email and a newsletter, which is two broadcasts and no sequence, and then conclude that email does not work for their category.

The size of the gap between those two approaches is visible in public data. Klaviyo's benchmarks, drawn from more than 183,000 of its customers, show triggered flows producing a 5.58% click rate against 1.69% for campaigns, with around 41% of total email revenue coming from 5.3% of sends. That is not an argument for sending more email. It is an argument for sending email that is caused by something.

Lifecycle emails that aren't a newsletter gives the five decisions, the trigger for each, the order to build them in, and the holdout that tells you whether any of it worked. The short version: build the message closest to the product's first value before you build the welcome.

At FrontRow I worked on the messages that went to new learners, and at Cintana Education's universities on the student journeys that ran for months rather than days. The length of the journey changes the trigger and nothing else about the method.

Can you see churn coming?

Yes, and earlier than most teams look. Churn announces itself as a series of small absences: a session count that drops below the account's own normal for three weeks running, a champion who stops logging in while their colleagues carry on, a support thread that ends without a reply, a card that expires.

The cheapest signal is the least interesting one. Recurly's analysis of 25 million subscription transactions from 2015, published in July 2016, found that about a fifth of all churn in its sample was involuntary, mostly failed payments from stale card details. A dunning sequence and a card-expiry warning are a week of work and they recover customers who never decided to leave.

The rest takes a scorecard, not a model. Churn you can see coming gives seven signals, the weights, the threshold, and the step almost everybody skips: scoring last quarter's churned accounts retroactively, as of 60 days before they canceled, so you know what your alert actually catches before you ask anyone to act on it.

What good looks like: one retention event everybody agrees on, a cohort curve that flattens and is redrawn monthly, a sequence of triggered messages with a holdout behind them, and a weekly list of accounts whose score dropped, with a named owner for each.